Statistical Models for Missing Data in Proteomic Studies of Gastric Cancer

Abstract

Disease diagnosis is often performed using a blood test for protein biomarkers which exhibit differential expression in diseased subjects as compared to healthy subjects. Discovery of new biomarkers enables cheaper and less invasive diagnosis. A method of biomarker discovery is the statistical analysis of proteomic mass spectrometry data to determine differences in protein concentration between groups of organisms. However, outcome-dependent missingness in proteomic mass spectrometry data hinders the extraction of useful information from the data and results in biased inference about these differences in protein expression. Existing methods of accounting for missing data, used for other, similar datasets such as those from RNA microarray experiments, assume missingness that is less severe and outcome-dependent than that which affects proteomic mass spectrometry data. These methods do not suffice to undo the bias, and new methods of statistical analysis are sought for biomarker discovery. We develop a joint statistical model for missing and observed data and apply it to a dataset from a gastric cancer experiment that has a large degree of outcome-dependent missingness in order to discover novel candidate biomarkers. A set of candidates is produced using the joint model. This set differs from the set of biomarker candidates produced in earlier work modelling the data without accounting for the outcome-dependent missingness.